# SpecGuard

## The problem An AI reviewer can sound certain even when the requirement is unclear. That creates a specific kind of review noise: clean changes get blocked by speculative risks, observations are presented as defects, and product decisions hide inside technical comments. The result is more review work, not less. A developer still has to work out which comments matter, what was actually tested, and whether the reviewer assumed behavior that nobody requested. ## What SpecGuard does SpecGuard adds a decision layer to Bob's review workflow. For each written requirement, it records one of four states: - supported by the cited implementation and checks; - contradicted by specific source or reproduced behavior; - needs a human decision because the intent is ambiguous; - unassessed because the available evidence is insufficient. A blocking finding must identify the requirement it contradicts. Test runs are stored as receipts. Source references are normalized and validated before the result reaches the dashboard.